Documentos de Académico
Documentos de Profesional
Documentos de Cultura
405
Abstract
The prominence of aviation security has increased dramatically in recent years.
As a reaction to the new threat situation large investments into modern security
technology have been made. State-of-the-art x-ray screening equipment provides
high resolution images and many image enhancement features. However, the
most expensive equipment is of limited value if the humans who operate it are
not selected and trained to perform their task accurately and efficiently. This
article presents results of studies conducted over the last four years which clearly
show that efficiency in x-ray screening can be increased substantially by
investing into human factors technologies that are based on results of visual
cognition, object recognition and psychophysics.
Keywords: Airport security, human factors, detection performance, computerbased training, pre-employment assessment
Introduction
406
Several image-based effects influence how well threat items can be recognized in
x-ray images (Schwaninger [2]; Schwaninger, Hardmeier, & Hofer [3]). As
illustrated in Fig. 1a, viewpoint can strongly affect recognition performance.
Systematic effects of viewpoint have been reported in many object recognition
studies (for recent reviews see Graf, Schwaninger, Wallraven & Blthoff [4];
Schwaninger [5]). In general, objects are difficult to recognize when depicted
from an unusual viewpoint and when diagnostic features are not visible. Object
recognition has also been shown to be dependent on exposure. When an object is
not similar to a stored view, it often becomes difficult to recognize.
a Viewpoint
b Superposition
c Bag
Complexity
Figure 1: Image-based factors: a) effect of viewpoint (canonical vs. noncanonical), b) superpositition by other objects (low vs. high), c) bag
complexity (low vs. high).
A second image-based factor is related to the fact that in x-ray images objects are
often superimposed by other objects. As depicted in Fig. 1b this effect of
superposition can impair detection performance substantially. A third factor is
bag complexity, which is determined by the number and type of other objects in
the bag (Fig. 1c). Since other objects distract attention, bag complexity can also
affect detection performance.
In order to measure how well people can cope with these image-based factors,
the x-ray Object Recognition Test (ORT) has been developed (Schwaninger [2]).
WIT Transactions on The Built Environment, Vol 82, 2005 WIT Press
www.witpress.com, ISSN 1743-3509 (on-line)
407
This test uses only guns and knives, object shapes that are known well by
novices. It contains 256 x-ray images, half of them with a threat item and half of
them with harmless bags. The guns and knives are depicted in easy and difficult
viewpoint, with little and much superposition in bags of low and high
complexity. All x-ray images of the ORT are in black-and-white, as color is
mainly diagnostic for the material of objects in the bag and thus is not very
helpful for novices. The ORT is fully computer-based. Each image is only
displayed for 4 seconds, simulating conditions of high passenger flow during
which average inspection times often are in the range of 3-5 seconds. For each xray image, participants have to decide whether the bag contains a threat item or
whether it is harmless.
Fig. 2 shows means and standard deviations from a study conducted with 284
candidates that applied for an x-ray screening job. Detection performance is
indicated by A, which can be calculated from hit and false alarm rates (for
details see Schwaninger, Hardmeier, & Hofer [3]; Hofer & Schwaninger [6]).
1.00
*
0.90
*
*
0.80
0.70
0.60
0.50
Unrotated
View
Rotated
View
Superposition Superposition
Low
High
Complexity
Low
Complexity
High
408
The previous section has shown that image-based factors such as view,
superposition and bag complexity affect how well threat items can be detected.
Even more important is the knowledge about which items are prohibited and
what they look like in x-ray images of passenger bags. Consider the three x-ray
images depicted in Fig. 3.
409
because it looks similar to a pen, a well known everyday object. The bag on the
right illustrates another problem that was already mentioned in the previous
section. When objects are depicted in an unusual view, they become difficult to
recognize if view-invariant diagnostic features are not available (for recent
reviews see Graf, Schwaninger, Wallraven & Blthoff [4]; Schwaninger [5]).
Since this is often the case in x-ray images, a large image library in which
objects are depicted in different views is necessary in order to achieve reliable
threat detection despite changes in viewpoint.
3.1 Multiple views TIP library
Essentially, object recognition means to match visual information from stimuli
(e.g. an x-ray image) to visual representations stored in memory. The content of
visual memory is dependent on exposure. Shapes are difficult to recognize if
they are not similar to objects that have been encountered earlier and were stored
in visual memory. This was illustrated in Fig. 3 at the beginning of section 3.
WIT Transactions on The Built Environment, Vol 82, 2005 WIT Press
www.witpress.com, ISSN 1743-3509 (on-line)
410
switchblade knife in Fig. 3b could be easily mistaken with a pen (see also Fig.
4c). A third very important but often neglected problem are effects of rotation.
As mentioned earlier, when objects are depicted in unusual viewpoints, they can
become very difficult to recognize. This was illustrated for a gun and scissors in
Fig. 3c. The upper row in Fig. 4d and Fig. 4e depicts these two threat items in
canonical view, i.e. the view that is easiest to recognize. As you can see in the
bottom row, effects of viewpoint can indeed be detrimental to recognition
performance.
D
i
e
n
s
t
w
a
f
f
e
S
I
G
2
2
5
ho00
ve45
ho45
ve85
ho85
vh45
c Each view in
3 planar rotations
411
becomes able to detect threat items reliably even if they are substantially rotated
away from the easiest view.
3.2 Individually adaptive training
After analysing the results of the first studies conducted at Zurich Airport in
2000 and 2001 we were looking for a CBT system that has a large image library
containing thousands of images depicting threat items in different viewpoints.
Moreover, such a system should be efficient and effective by applying
individually adaptive training algorithms that calculate for each screener which
items and views need to be trained particularly. The results of our first survey
were that all existing CBT systems had only small image libraries and they did
not feature individually adaptive training algorithms. Instead, the commercially
available systems focussed on providing many image processing features such as
negative image, black and white, edge detection etc. It actually sometimes
seemed that the quality of a CBT would be dependent on providing an exact
replica of the x-ray machine with as many image processing and enhancement
features as possible. However, the goal of the training system we had in mind
was to train screeners to detect threat items reliably within the first 3-5 seconds
of inspection time before image processing features might be applied. In order to
achieve this, a scientifically based training system was needed that features a
large image library, sophisticated adaptive training algorithms and feedback
mechanisms that are based on results of perceptual learning. Since nothing like
that existed, we decided to create such a system on our own. In a close
collaboration between psychologists, vision and computer scientists as well as
aviation security experts the first version of X-Ray Tutor was built in 2002 and
installed at Zurich airport. It is now used by some of the largest security
organizations in the world. The United States Transportation Security
Administration is using X-Ray Tutor at airports across the country since spring
2004 as part of its recurrent training and professional development program for
airport security screeners. The Canadian Air Transport Security Authority has
started using X-Ray Tutor at several airports as part of a research project with the
University of Zurich on x-ray image tutoring technology for aviation security
screeners. X-Ray Tutor is operational in Switzerland since 2002, at 17 airports in
Germany since 2003, and since 2004 in Belgium and other European countries.
At the heart of X-Ray Tutor are results from scientific studies on how the human
brain processes visual information in order to recognize objects from different
viewpoints (Schwaninger [7], [8]; for recent overviews on object recognition see
Graf, Schwaninger, Wallraven & Blthoff [4]; Schwaninger [5]). The system
uses a sophisticated algorithm to create training sessions for each individual
student based on detection performance, response time and x-ray image
difficulty ratings. The individually adaptive algorithm of X-Ray Tutor starts with
easy x-ray images. Then, view difficulty is increased and trainees learn to detect
threat items in different viewpoints. X-Ray Tutor uses the view-based image
library described above, which allows presenting each threat item in 72 different
viewpoints. As mentioned before, it is not necessary to show each item in all
viewpoints. The adaptive algorithm of X-Ray Tutor shows for each trainee
WIT Transactions on The Built Environment, Vol 82, 2005 WIT Press
www.witpress.com, ISSN 1743-3509 (on-line)
412
413
reports than can be imported easily into applications for further analysis such as
standard office software or analysis tools for large databases. Using the reporting
feature, detection performance can be analyzed and visualized across time, for
individual screeners, groups of screeners, specific FTIs and groups of FTIs. This
data can then be used in order to answer several important questions such as for
example:
- How does detection performance increase over time?
- Which threat items are difficult to recognize?
- How does viewpoint affect threat detection performance?
- How reliable are individual performance measurements?
These questions are currently being investigated in research projects conducted
in our lab. The following section summarizes the results of a study that was
conducted in 2002/2003 in order to measure training effectiveness and
investigate perceptual learning mechanisms. A detailed description of this study
can be found in Schwaninger and Hofer [9].
3.3 Measuring Effectiveness and investigating perceptual learning
The study was conducted in the period of December 2002 May 2003 with 72
participants. None of them had received computer-based training before. Each
week, 1-2 training sessions of 20 minutes were completed during the six months
period. X-Ray Tutor HBS for training detection of improvised explosive devices
(IEDs) was used. The research questions were as follows:
- Does IED detection performance increase over time as a result of individually
adaptive training?
- Are there large differences in detection performance when images are displayed
4 vs. 8 seconds?
- Does training result in a decrease of response time? If so, is that decrease
different for threat images vs. harmless bag images?
Four tests were conducted in which new IEDs were used that had not been
shown previously during training. Display durations at test were 4 and 8 seconds
(during training, images were displayed for a maximum of 12 seconds). The four
tests were of equal difficulty as determined in a pilot study. Signal detection
theory was used to analyze the data (Green & Swets [10]). Detection
performance was measured using d = z(H) z(FA), whereas H represents the hit
rate, FA the false alarm rate and z the z-transformation. Similar results were
obtained using non-parametric A instead of d as performance measure (for a
comprehensive introduction on signal detection and x-ray screening see
Schwaninger [11]; for more detailed information on different detection measures
of x-ray detection performance see Hofer and Schwaninger [6]). The results of
the four tests are displayed in Fig 7. There was only a small effect of display
duration, i.e. detection performance was similar for display durations of 4 and 8
seconds. This is consistent with the view that recognition is a fast process
occurring within the first few seconds of image inspection. More important is the
knowledge about threat items, which is dependent on training. Indeed, large
WIT Transactions on The Built Environment, Vol 82, 2005 WIT Press
www.witpress.com, ISSN 1743-3509 (on-line)
414
4 sec
8 sec
80
60
40
4 sec
8 sec
20
0
1
Dec/Jan
Feb/Mar
Dec/Jan
Mar/April April/May
Feb/Mar
Mar/April April/May
Non-Threat Images
4 sec
8 sec
6
5
4
3
2
4 sec
8 sec
6
5
4
3
2
Dec/Jan
Feb/Mar
Mar/April
April/May
Dec/Jan
Feb/Mar
Mar/April
April/May
415
There was also an effect of display duration, although much smaller in effect size
than the overall training effect. For harmless bags average response times
remained constant at about 5 seconds. This finding is consistent with the
assumption that individually adaptive CBT increases the number and strength of
representations in visual memory. Because the participants have learnt what
IEDs look like they became able to detect them reliably within a few seconds of
inspection time. However, the acquired knowledge does not help if there is no
threat item in the bag. Since no substantial reduction in response time was
observed for harmless bag images, the learning effect seems rather related to
visual memory representations than to increased general visual processing
capacities.
Conclusion
Acknowledgements
I am thankful to airport and police authorities who provided data and funding for
conducting the studies reported in this paper. Thanks to Franziska Hofer and
Judith Riegelnig for their help in preparing the manuscript.
References
[1] Blthoff, H.H., Lee, S.-W., Poggio, A., & Wallraven, C., Biologically
motivated computer vision, Lecture Notes in Computer Science, 2525,
Berlin: Springer, 2002.
[2] Schwaninger, A., Evaluation and selection of airport security screeners.
AIRPORT, 02/2003, 14-15, 2003.
[3] Schwaninger, A., Hardmeier, D., & Hofer, F., Measuring visual abilities and
visual knowledge of aviation security screeners. IEEE ICCST Proceedings,
38, 258-264, 2004.
[4] Graf, M., Schwaninger, A., Wallraven, C., & Blthoff, H.H.,
Psychophysical results from experiments on recognition & categorisation.
Information Society Technologies (IST) programme, Cognitive Vision
Systems - CogVis (IST-2000-29375), 2002.
[5] Schwaninger, A., Object recognition and signal detection. In: B. Kersten &
M.T. Groner (Eds.), Praxisfelder der Wahrnehmungspsychologie. Bern:
Huber, 2004.
WIT Transactions on The Built Environment, Vol 82, 2005 WIT Press
www.witpress.com, ISSN 1743-3509 (on-line)
416
[6] Hofer, F., & Schwaninger, A., Reliable and valid measures of threat
detection performance in x-ray screening, IEEE ICCST Proceedings, 38,
303-308, 2004.
[7] Schwaninger, A., Training of airport security screeners. AIRPORT, 05/2003,
11-13, 2003.
[8] Schwaninger, A., Computer based training: a powerful tool to the
enhancement of human factors. Aviation Security International, FEB/2004,
31-36, 2004.
[9] Schwaninger, A. & Hofer, F. Evaluation of CBT for increasing threat
detection performance in x-ray screening. In: K. Morgan and M. J. Spector,
The Internet Society 2004, Advances in Learning, Commerce and Security
(pp. 147-156). Wessex: WIT Press, 2004.
[10] Green, D. M. & Swets, J. A., Signal detection theory and psychophysics,
Wiley: New York, 1996.
[11] Schwaninger, A., Reliable measurements of threat detection. AIRPORT,
01/2003, 22-23, 2003.
WIT Transactions on The Built Environment, Vol 82, 2005 WIT Press
www.witpress.com, ISSN 1743-3509 (on-line)